Model Reference
Model Reference
Last verified: October 2026. Every Ollama tag and download size below was checked against its page on ollama.com. New releases land weekly, so if something looks outdated, check Ollama’s library and HuggingFace trending.
How the RAM numbers work. “Download” is the size Ollama shows on the model’s tags page. “RAM” is that size plus about 4 GB for the OS, the runtime, and a at everyday context lengths, rounded up. Long contexts need more. Close to your limit, pick the next size down.
Recommended Starters
One pick per tier. Install Ollama, then run the command.
| Your RAM | Model | Command | Why this one |
|---|---|---|---|
| 4 GB or less | Qwen 3.5 0.8B | ollama pull qwen3.5:0.8b | 1.2 GB download. Tiny but functional, and tight on a 4 GB machine. |
| 8 GB | Qwen 3.5 4B | ollama pull qwen3.5:4b | with . 3.3 GB download, 256K . |
| 16 GB | Qwen 3.5 9B | ollama pull qwen3.5:9b | Same family with thinking mode and image input. 6.6 GB download, 256K context. Gemma 4 12B (gemma4:12b) is the other good pick here. |
| 32 GB | Qwen 3.8 27B | ollama pull qwen3.8:27b | Dense 27B with thinking mode and image input. 18 GB download, 256K context, Apache 2.0. |
| 64 GB | Qwen 3.8 27B at Q8 | ollama pull qwen3.8:27b-q8_0 | The same model at higher precision (30 GB download). For more speed, try the qwen3.6:35b, which has only 3B active . |
| 96 GB+ | GPT-OSS 120B | ollama pull gpt-oss:120b | 65 GB download, 128K context, Apache 2.0. Needs about 96 GB to run comfortably. |
All Models by Hardware Tier
8 GB RAM: Small Models (up to 4B)
| Model | Params | Download | RAM | Strengths | Ollama |
|---|---|---|---|---|---|
| Qwen 3.5 4B | 4B (dense) | 3.3 GB | ~7 GB | Image input, thinking mode, 256K context | qwen3.5:4b |
| Nemotron 3 Nano 4B | 4B | 2.8 GB | ~7 GB | Hybrid Mamba-, tool calling, thinking mode, 256K context | nemotron-3-nano:4b |
| Phi-4 Mini | 3.8B (dense) | 2.5 GB | ~7 GB | Microsoft, MIT license, tool calling, 128K context | phi4-mini |
| Gemma 4 E2B | ~2B effective | 4.6 GB | ~9 GB | Image and audio input, 128K context. Tight on 8 GB, use short contexts. | gemma4:e2b |
16 GB RAM: Medium Models (7B-14B)
| Model | Params | Download | RAM | Strengths | Ollama |
|---|---|---|---|---|---|
| Qwen 3.5 9B | 9B (dense) | 6.6 GB | ~11 GB | Image input, thinking mode, 256K context | qwen3.5:9b |
| Gemma 4 E4B | ~4B effective | 6.6 GB | ~11 GB | Image and audio input, 128K context | gemma4:e4b |
| Gemma 4 12B | 12B (dense) | 7.7 GB | ~12 GB | Apache 2.0, 256K context | gemma4:12b |
| Phi-4 14B | 14B (dense) | 9.1 GB | ~13 GB | Strong reasoning for its size. Only 16K context. | phi4:14b |
32 GB RAM: Large Models (20B-35B)
| Model | Params | Active | Download | RAM | Strengths | Ollama |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B | 27B | 27B (dense) | 18 GB | ~22 GB | Thinking mode, image input, 256K context, Apache 2.0 | qwen3.8:27b |
| Qwen 3.6 35B-A3B | 35B | 3B (MoE) | 23 GB | ~27 GB | Fast for its size, 256K context, Apache 2.0 | qwen3.6:35b |
| Gemma 4 31B | 31B | 31B (dense) | 19 GB | ~23 GB | 256K context, Apache 2.0 | gemma4:31b |
| Gemma 4 26B-A4B | 26B | ~4B (MoE) | 16 GB | ~20 GB | Lighter than the 31B, 256K context | gemma4:26b |
| GPT-OSS 20B | 21B | 3.6B (MoE) | 14 GB | ~18 GB | OpenAI’s model, 128K context, Apache 2.0 | gpt-oss:20b |
| Muse Glimmer | 30B | n/a | 18 GB | ~22 GB | Meta, built for agent tasks, image input, 128K context, Apache 2.0 | muse-glimmer |
| Nemotron 3.5 Lightning | 30B | 3B (MoE) | 19 GB | ~23 GB | NVIDIA, 1M context, tool calling. Long contexts need much more RAM. | nemotron-3.5-lightning |
| GLM-4.7 Flash | 30B | 3B (MoE) | 19 GB | ~23 GB | Agentic coding, tool calling, 198K context | glm-4.7-flash |
96 GB+ RAM: Very Large Models (100B+)
These do not fit a 64 GB machine at the default quantization.
| Model | Params | Active | Download | RAM | Strengths | Ollama |
|---|---|---|---|---|---|---|
| GPT-OSS 120B | 117B | 5.1B (MoE) | 65 GB | ~69 GB | 128K context, Apache 2.0 | gpt-oss:120b |
| Llama 4 Scout | 109B | 17B (MoE) | 67 GB | ~71 GB | Image input, 10M token context per Ollama’s listing | llama4:16x17b |
| Nemotron 3.5 Super | 120B | 12B (MoE) | 80 GB | ~84 GB | 1M context, tool calling | nemotron-3.5-super |
| Qwen 3.5 122B-A10B | 122B | 10B (MoE) | 81 GB | ~85 GB | 256K context, Apache 2.0 | qwen3.5:122b-a10b |
| Nemotron 3 Super | 120B | 12B (MoE) | 87 GB | ~91 GB | 256K context, tool calling | nemotron-3-super |
Server / Multi-GPU: Frontier-Class Open Models
These need hundreds of GB of memory or a multi-GPU server. The newest ones run on Ollama only as cloud models (:cloud tags), so for local use you need the Hugging Face weights and a serving stack like vLLM or llama.cpp.
| Model | Total params | License | Where |
|---|---|---|---|
| Llama 4 Maverick (17B active) | 400B | Llama 4 Community License | Ollama llama4:128x17b, 245 GB download, 1M context |
| DeepSeek V4.1 Flash | 763B | MIT | HuggingFace, 1M context. Ollama: cloud only. |
| DeepSeek V4 Pro | 1.6T | MIT | HuggingFace. Ollama: cloud only. |
| GLM-5.3 | 753B | Custom (“other”) | HuggingFace, 1M context. Ollama: cloud only. |
| Kimi K3 | 2.78T | Custom (“other”) | HuggingFace, 1M context. Ollama: cloud only. |
| Qwen 3.8 2.4T-A95B (95B active) | 2.4T | Custom (“other”) | HuggingFace |
| Mistral Small 4 | 119B | Apache 2.0 | HuggingFace |
Models by Use Case
These are editorial picks from the tables above, not a benchmark ranking.
Best for Coding
| Tier | Model | Why |
|---|---|---|
| Small (8 GB) | Qwen 3.5 4B | Thinking mode helps with code reasoning |
| Medium (16 GB) | Qwen 3.5 9B | Thinking mode, 6.6 GB download |
| Large (32 GB) | Qwen 3.8 27B | Model card reports 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench v6 (vendor-reported) |
| XL (96 GB+) | GPT-OSS 120B | Strong agentic coding with only 5.1B active params |
| Server | DeepSeek V4 Pro / Kimi K3 / GLM-5.3 | Current open-weight flagships |
Best for Reasoning and Math
| Tier | Model | Why |
|---|---|---|
| Small (8 GB) | Qwen 3.5 4B | Thinking mode |
| Medium (16 GB) | Qwen 3.5 9B | Thinking mode, 6.6 GB download |
| Large (32 GB) | Qwen 3.8 27B | Model card reports 89.2 on GPQA Diamond (vendor-reported) |
| XL (96 GB+) | Qwen 3.5 122B-A10B | Thinking mode with 10B active params |
| Server | DeepSeek V4 Pro / Qwen 3.8 2.4T-A95B | Largest open checkpoints |
Best for Long Context
| Tier | Model | Context | Why |
|---|---|---|---|
| Small (8 GB) | Nemotron 3 Nano 4B | 256K tokens | Hybrid Mamba-Transformer, 2.8 GB download |
| Medium (16 GB) | Gemma 4 12B | 256K tokens | 7.7 GB download |
| Large (32 GB) | Nemotron 3.5 Lightning | 1M tokens | Needs far more than 23 GB at full context |
| XL (96 GB+) | Llama 4 Scout | 10M tokens | Ollama’s listed context. Real usable length depends on your KV cache memory. |
Best for Multimodal (Text + Images)
| Tier | Model | Modalities |
|---|---|---|
| Small (8 GB) | Qwen 3.5 4B | Text + images |
| Small (8 GB) | Gemma 4 E2B | Text + images + audio |
| Medium (16 GB) | Qwen 3.5 9B | Text + images |
| Medium (16 GB) | Gemma 4 E4B | Text + images + audio |
| Large (32 GB) | Qwen 3.8 27B | Text + images |
| XL (96 GB+) | Llama 4 Scout | Text + images |
Best for Writing and Creative Tasks
I have no benchmark for this one. These are picks based on size and context length, so test them on your own prompts.
| Tier | Model | Why |
|---|---|---|
| Medium (16 GB) | Gemma 4 12B | Newest mid-size general model |
| Large (32 GB) | Qwen 3.8 27B | Large context for keeping tone across long pieces |
| XL (96 GB+) | GPT-OSS 120B | Large general model that fits a 128 GB machine |
Instruct vs. Base Models
Base models predict the next token and are not trained to be helpful. They will continue a prompt as if completing a document, not answer a question.
Instruct models are fine-tuned to follow instructions and hold conversations. For everyday use, always download the instruct variant.
When downloading, look for “Instruct” or “it” in the model name (e.g., gemma-4-12B-it). HuggingFace model cards label these clearly. The only reason to use a base model is if you are doing further fine-tuning and want the raw weights.
Most models in the Ollama library are instruct variants by default unless the name says otherwise.
Model Families: Who Makes What
Qwen (Alibaba)
- Qwen 3.5: February 2026. Sizes 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B. Apache 2.0.
- Qwen 3.6: April 2026. 27B and 35B-A3B on Ollama. Apache 2.0.
- Qwen 3.8: August 2026. 27B is Apache 2.0, 256K native context (extendable to 1M), thinking mode on by default, image input. The 2.4T-A95B flagship uses a custom license.
- Links: HuggingFace Qwen3.8-27B · Ollama qwen3.8 · Ollama qwen3.6 · Ollama qwen3.5
GPT-OSS (OpenAI)
- Released: August 2025
- Sizes: 20B (3.6B active), 120B (5.1B active)
- License: Apache 2.0
- Key features: MoE, MXFP4 weights, tool calling, thinking mode, 128K context
- Links: HuggingFace · GitHub · Ollama
Nemotron (NVIDIA)
- Nemotron 3: Nano 4B, Nano 30B-A3B (December 2025), Super 120B-A12B. NVIDIA Open Model License.
- Nemotron 3.5: Lightning 30B-A3B (August 2026, 1M context, OpenMDW 1.1 license) and Super 120B-A12B (1M context).
- Key features: Hybrid Mamba-Transformer MoE, tool calling, agentic reasoning
- Links: NVIDIA Developer · Ollama Lightning · Ollama 3.5 Super · Ollama Nano
Llama 4 and Muse (Meta)
- Llama 4 released: April 2025
- Sizes: Scout 109B (17B active), Maverick 400B (17B active)
- License: Llama 4 Community License
- Key features: MoE, image input, Scout has a 10M token context window on Ollama
- Muse Glimmer: Meta’s newer local model. 30B, Apache 2.0, 128K context, 18 GB on Ollama.
- Links: llama.com · Ollama Llama 4 · Ollama Muse Glimmer
Gemma 4 (Google DeepMind)
- Released: April 2026 (12B added June 2026)
- Sizes: E2B, E4B, 12B, 26B-A4B, 31B
- License: Apache 2.0
- Key features: E2B and E4B take image and audio input and have 128K context. 12B, 26B and 31B have 256K context.
- Links: DeepMind Blog · HuggingFace Blog · Ollama
GLM (Zhipu AI / Z.AI)
- Local pick: GLM-4.7 Flash (30B-A3B, MIT, January 2026), 19 GB on Ollama, 198K context.
- Flagship: GLM-5.3 (753B, August 2026, custom license). Ollama cloud only.
- Links: Ollama GLM-4.7 Flash · HuggingFace GLM-5.3
DeepSeek
- Current: V4 Pro (1.6T) and V4.1 Flash (763B, September 2026), both MIT with 1M context on Flash. Neither is local-size.
- Older: R1 and its distilled sizes (1.5B to 70B) are still on Ollama as
deepseek-r1. - Links: HuggingFace V4.1 Flash · Ollama R1
Kimi (Moonshot AI)
- Current: Kimi K3 (June 2026), 2.78T total params, custom license, 1M context. Ollama cloud only.
- Links: HuggingFace
Mistral
- Mistral Small 3: 24B, January 2025, Apache 2.0, 14 GB on Ollama (
mistral-small), 32K context. It still works but is no longer a first pick. - Mistral Small 4: 119B, Apache 2.0, March 2026. Server-class.
- Links: Ollama · HuggingFace Small 4
Phi-4 (Microsoft)
- Released: 2025
- Sizes: Mini 3.8B (128K context), 14B (16K context)
- License: MIT
- Links: Ollama Phi-4 Mini · Ollama Phi-4
Where to Find Models
- Unsloth Dynamic GGUFs produces dynamic quantizations that are widely considered the best GGUF variants for most models. The dynamic format allocates more bits to sensitive layers and fewer to layers where precision matters less, giving better quality at the same file size. Start here when downloading GGUFs from HuggingFace.
- Ollama Library has pre- models with one-command install. Best for getting started quickly.
- HuggingFace Models is the largest collection. Original weights plus community quantizations (, GPTQ, AWQ).
- LM Studio provides a visual model browser with one-click downloads from HuggingFace.
Understanding Model Names
When you see a model name like qwen3.5:9b-q4_K_M, here’s what each part means:
| Part | Meaning | Example |
|---|---|---|
| Family | Who made it and which generation | qwen3.5 = Alibaba’s Qwen, version 3.5 |
| Size | Total count | 9b = 9 billion parameters |
| Active (if MoE) | Parameters used per token | A3B = 3 billion active |
| Quantization | Compression level | q4_K_M = 4-bit, , medium quality |
quick guide:
| Level | Quality | Notes |
|---|---|---|
| Q8 | ~99% of original | Largest file. Near-lossless. |
| Q6_K | ~97% | Slightly smaller, barely any loss. |
| Q5_K_M | ~95% | Good default if you have the space. |
| Q4_K_M | ~93% | The sweet spot for most people. Half the size. |
| Q3_K | ~85% | Noticeable quality drop. Only if you’re tight on memory. |
| Q2_K | ~70% | Real quality loss. Last resort. |
Want to understand how quantization works? See Module 2: Compressed Models for the intuition, or Module 3B: How Models Fit for the deep dive.
Emerging Techniques Changing the Game
TurboQuant (Google)
Google Research’s March 2026 blog post describes compressing KV caches to 3 bits without retraining. The authors report at least 6x lower KV memory in needle-in-a-haystack tests and up to 8x faster attention logits on H100 GPUs (4-bit keys versus 32-bit). Those are the authors’ own results. Support in llama.cpp, MLX and Ollama varies, so check their release notes before counting on it. The paper is arXiv 2504.19874.
LLM in a Flash (Apple)
Runs models larger than your available by streaming weights from SSD/flash storage. Apple Research describes the technique in this paper. It is slow compared with keeping weights in memory.
Mixture of Experts (MoE) Everywhere
Most new large models use MoE: huge total parameter counts but only a fraction active per token. A “120B” MoE model with 5B active runs at roughly 5B speed and compute cost, but stores knowledge across all 120B parameters. You need RAM for the full 120B, which is why the tables above show models larger than you might expect for a given tier.
For the latest in model releases and AI news, see our News Feed and Research Feed.